1 / 30

Distributed Algorithmic Mechanism Design for Network Problems

Distributed Algorithmic Mechanism Design for Network Problems. Rahul Sami. Ph.D. Dissertation Defense. Advisor: Joan Feigenbaum. Committee: Ravindran Kannan Arvind Krishnamurthy Scott Shenker (ICSI & Berkeley).

yon
Download Presentation

Distributed Algorithmic Mechanism Design for Network Problems

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Distributed Algorithmic Mechanism Design for Network Problems Rahul Sami Ph.D. Dissertation Defense Advisor: Joan Feigenbaum Committee: Ravindran Kannan Arvind Krishnamurthy Scott Shenker (ICSI & Berkeley)

  2. Mechanism Design Framework Private “types” Strategies Mechanism t1 a1 p1 Agent 1 Output O tn an pn Agent n Payments Agent i chooses strategy aito maximize her welfare.

  3. Strategyproof Mechanisms Strategyproof mechanism: Regardless of what other agents do, each agent i maximizes her welfare by revealing her true private type information ti. Group-strategyproof mechanism: Even if agents can collude, no group of agents can benefit from not revealing their true types– regardless of the strategies of non-members, any strategy that makes one group member strictly better off would make at least one group member strictly worse off.

  4. Algorithmic Mechanism Design Algorithmic Mechanism Design (Nisan-Ronen ’01): • Introduced computational efficiency into mechanism-design framework. Polynomial-time computable O( ) and pi( ) Centralized model of computation Distributed AMD (Feigenbaum-Papadimitriou-Shenker ’01) Computation is distributed across the network. • “Good network complexity”: −Polynomial-time local computation −Modest in total number of messages, messages per link, and maximum message size

  5. Thesis Statement “The distributed-computing context can have a major impact on the feasibility of a mechanism.” Thesis supported with results on • Multicast Cost-Sharing • Interdomain Routing

  6. Multicast Cost SharingMechanism-design Problem Receiver Set Source Which usersreceive the multicast? 3 3 1,2 3,0 1 5 5 2 Cost Shares How much does each receiver pay? 10 6,7 1,2 Users’ utilities Link costs

  7. Earlier Work on Multicast Cost Sharing Moulin and Shenker ‘01 Two mechanisms: • Efficient: marginal cost (MC) • Budget-balanced: Shapleyvalue (SH) Feigenbaum, Papadimitriou and Shenker ‘01 • Distributed algorithm for MC with good network complexity • Restricted lower bound on network complexity of SH

  8. Multicast Cost-Sharing Results (1) [ Feigenbaum-Krishnamurthy-Sami-Shenker. To appear in Theoretical Computer Science. Extended abstract appeared in FSTTCS ’02. ] • Any deterministic or randomized algorithm for any exactly budget-balancedgroup-strategyproof mechanism must send (n) bits on a single link in the worst case, where n is the number of users. • Any algorithm to compute a κ-approximately budget-balanced group-strategyproof mechanism, for κ< 2, must send (log n / log κ) bits on a single link in the worst case. • There is no mechanism that is strategyproof, approximately budget-balanced, and approximately efficient.

  9. Multicast Cost-Sharing Results (2) [ Archer-Feigenbaum-Krishnamurthy-Sami-Shenker, to appear in Games and Economic Behavior. ] Constructed a mechanism SSF (Scaled Step-Function) that is • Group-strategyproof • Network complexity: log n ( ) O values communicated on a link log κ Revenue 1 1 • Budget-deficit bound:   Cost κh • Efficiency bound: Eff.(SSF)  Eff.(SH) – (κh-1)U κ> 1 controls trade-off of computational goalsvs.economic goals.

  10. UUNET Qwest WorldNet Sprint Mechanism Design for Interdomain Routing Agents: Autonomous Systems (ASs) Two formulations: Lowest-cost Routing General Policy Routing Routes currently computed with the Border Gateway Protocol (BGP).

  11. Agents’ types: Per-packet costs {ck} Outputs: {route(i, j)} Lowest-Cost Routing Payments: {pk} Objectives: Lowest-cost paths (LCPs) Strategyproofness “BGP-based” distributed algorithm Prior work: [Nisan-Ronen ’01]: Links (not nodes) are strategic agents. Strategyproof mechanism for single source-destination pair. Polynomial-time centralized computation. [Hershberger-Suri ’01]: Faster payment computation for Nisan-Ronen mechanism.

  12. Lowest-Cost Routing Results [ Feigenbaum-Papadimitriou-Sami-Shenker, PODC ‘02]: • For a biconnected network, if LCP routes are always chosen, there is a unique strategyproof mechanism that gives no payment to nodes that carry no transit traffic. • Prices required by this mechanism can be computed by a “BGP-based” algorithm with routing tables of size O(nd) (a constant factor increase over BGP) and convergence time at most max(d,d’) stages of computation, where d = maxi,j(#hops in LCP from i to j ) d’ = maxi,j,k(# hops in lowest-cost i-j path not using k )

  13. General Policy-Routing Mechanism Design Problem Per-packet ck is an unrealistic cost model. • AS route preferences are influenced by reliability, customer-provider relationships, etc. General Policy Routing: − For all i,j , AS i assigns avalue ui(Pij)to each potentialroute Pij. − Mechanism-Design Goals: • Maximize W = ui(Pij). • For each destination j, {Pij} forms a tree. • Strategyproofness • BGP-based distributed algorithm i,j

  14. Policy-Routing Results [ Feigenbaum-Griffin-Ramachandran-Sami-Shenker ]: Arbitrary valuation functions ui(Pij) • NP-hard to find routing tree that maximizes W. • NP-hard even to approximate maximum W within O(n1/4 - ). Next-hop preferences: ui(Pij)depends only on next-hop ASa. • Captures preferences due to customer/provider/peer agreements. • There is a unique strategyproof mechanism that maximizes W and does not pay non-transit nodes. • This mechanism is polynomial-time computable (in a centralized model). • BGP-based distributed algorithm is infeasible: • May take (n) stages to converge. • May require (n) messages for every preference change.

  15. Other Research done at Yale Theoretical Computer Science [ Feigenbaum-Fortnow-Pennock-Sami, EC ‘03]: Computation in a Distributed Information Market [ Batu-Ergun-Kilian-Magen-Rubinfeld-Raskhodnikova-Sami, STOC ‘03]: A Sublinear Algorithm for Weakly Approximating Edit Distance Computer Architecture [ Henry-Kuszmaul-Loh-Sami, ISCA ‘00]: Circuits for Wide-Window Superscalar Processors [ Henry-Loh-Sami, ISHPC ‘02]: Speculative Clustered Caches for Clustered Processors [ Loh-Sami-Friendly, WDDD ‘02]: Memory Bypassing: Not Worth the Effort

  16. General Policy-Routing Problem Statement Consider each destination j separately. Each AS i assigns avalue ui(Pij)to each potentialroute Pij. a j i b d Mechanism-design goals: • Maximize W = u i(Pij). • For each destination j, {Pij} forms a tree. • Strategyproofness • BGP-based distributed algorithm i

  17. NP-Hardness with Arbitrary Valuations Approximability-preserving reduction from Independent-set problem: j f b e2 e2 e2 a f b a e1 e1 e1 e2 b e1 ta tb tf Paths from terminals ta, tb, tf have valuation 1, all other paths 0. NP-hard to compute maximumWexactly. NP-hard to compute O(n1/4 - ) approximation to maximumW.

  18. Next-Hop Preferences • ui(Pij)depends only on next-hop ASa. • Captures preferences due to customer/provider/peer agreements. For each destination j , optimal routing tree is a Maximum-weight Directed Spanning Tree (MDST): a Edge weight = ui([i a … j]) c i j b

  19. Strategyproof Mechanism Let T* = Maximum weight directed spanning tree (MDST) in G T-i = MDST in G – {i} • For 2-connected networks, there is a unique strategyproof mechanism that always picks a welfare-maximizing routing tree and never pays non-transit nodes. The payments required for this mechanism are pi = W(T*) – ui(T*) – W(T-i ) • Routes and payments can be computed in polynomial time (in a centralized computational model.)

  20. BGP-based Computational Model (1) Follow abstract BGP model of Griffin and Wilfong: Network is a graph with nodes corresponding to ASes and bidirectional links; intradomain-routing issues and route aggregation are ignored. Each AS has a routing table with LCPs to all other nodes: Cost/Pref. AS Path Dest. AS1 AS3 AS5 AS1 3 AS2 AS7 AS2 2 Entire paths are stored, not just next hop.

  21. Advertise modified routes Receive routes from neighbors Update routing table BGP-based Computational Model (2) • An AS “advertises” its routes to its neighbors in the AS graph, whenever its routing table changes. • The computation of a single node is an infinite sequence of stages: • Complexity measures: − Number of stages required for convergence − Total communication

  22. Proving Hardness for “BGP-based” Routing Algorithms Requirements for routing to any destination j : [P1] Each AS uses O(l) space for a route of lengthl. [P2] The routes converge in O(log n) stages. [P3] Most changes should not have to be broadcast to most nodes: i.e., there are o(n) nodes that can trigger (n) updates when they fail or change valuation. Hardness results must hold: • For “Internet-like” graphs: O(1) average degree O(log n) diameter • For an open set of preference values in a small range.

  23. Long Convergence Time of MDST Theorem 1: An algorithm for MDST mechanism cannot satisfy property [P2]. Proof sketch: 1 1 1 1 1 1 1 2 2 2 2 • Network has diameter O(log n), but MDST has a path of length (n). • BGP routes in a hop-by-hop manner  (n) stages required for convergence.

  24. Extensive Dynamic Communication Theorem 2: A distributed algorithm for the MDST mechanism cannot satisfy property [P3]. Proof outline: (i) Construct a network and valuations such that for (n) nodes i, T-i is disjoint from the MDST T*. • A change in the valuation of any node a may change pi = W(T*) – ui(T*) – W(T-i). (iii) Node i (or whichever node stores pi) must receive an update when this change happens.  (n) nodes can each trigger (n) update messages.

  25. L- 2k -1 B R B R L- 2k -1 k-cluster k-cluster (k+1)-cluster Network Construction (1) (a) Construct 1-cluster with two nodes: L-1 red port blue port L  2 log n + 4 B R L-1 1-cluster (b) Recursively construct (k+1)-clusters: red port blue port

  26. 5 7 9 9 9 9 B R B B R R B R 9 9 9 9 7 7 5 3 2 j Network Construction (2) (c) Top level: m-cluster with n = 2m + 1 nodes. blue port red port B R m-cluster L- 2m -1 L- 2m -2 j Destination Final network (m = 3):

  27. Optimal Spanning Trees Lemma: W(blue tree) = W(red tree) + 1 W(any other sp.tree) + 2 Proof: If a directed spanning tree has red and blue edges, we can increase its weight by at least 2: L- 2k -3 L- 2k -3 B R k-cluster k-cluster (k+1)-cluster L- 2k -1 L- 2k -3 B B k-cluster k-cluster (k+1)-cluster

  28. 5 7 9 9 9 9 B R B B R R B R 9 9 9 9 7 7 5 3 2 j Proof of Theorem 2 MDST T* is the blue spanning tree. For any blue nodeB, T-B is the red spanning tree on N–{B}. A small change in any edge, red or blue, changes pB = W(T*) – uB(T*) – W(T-B)  Any change triggers update messages to all blue nodes!

  29. Summary Dissertation shows that the distributed-computing context can have a major impact on mechanism design. Multicast cost-sharing • Budget-balanced, group-strategyproof mechanisms are hard. • Tractable mechanism with bounded budget deficit. Interdomain routing • Lowest-cost routing mechanism is easy in BGP-based computational model. • Next-hop policy routing mechanism is hard in BGP-based computational model.

  30. Some Open Questions.. • Multicast cost-sharing: Can we achieve constant-factor approximate budget-balance with O(log n) messages per link? • Routing: Can we prevent strategy in computation without digital signatures? • General: • New solution concepts • Other problem domains (caching, wireless, ...)

More Related